Repository navigation
fix(native-chat): a resting chat lists its model the way the running agent does, so no stray effort picker appears - #24267
Conversation
…ild does, with no phantom effort
A chat at rest (no live child) answered its options from the host's model
catalog alone. When the chat's model is one the catalog does not list, the
answer left it out, the client filled the gap from its static seed, and the
composer showed an effort control ("Medium") that the live child never offers;
the control vanished again once the child was back.
The live Claude and Codex answers and the resting answer now build their model
list through one function: the catalog's rows, plus the current model when the
catalog does not list it, with no effort levels. Live and rest can no longer
disagree about it.
…e child would, so a listed model keeps its effort When the host has no model catalog for the account yet (the first read before its probe returns, a failing probe, an account-home lookup that failed, or a WSL-pinned chat no live child has written one for), the resting answer listed only the saved model, as unlisted, and a listed model such as opus lost its effort control. It now rests on the list a running child falls back to: Claude's built-in models, then the shared function. Codex lists nothing without its catalog, so the resting answer leaves the list to the client's own defaults, as a failed live read does.
… model, as before With no catalog, the built-in list's default model is a guess, not the account's: naming it moved a chat the running agent reported on Opus to Sonnet. Only a real listing names the default now; otherwise the answer names none and the client keeps the model the agent last reported. The comment on Codex's no-catalog answer now says what it does: the client fills the current model from its own defaults, unchanged from before.
…ions-at-rest-match-live
Review status: ready for review (must land before #24235)The problem. When a native Claude or Codex chat is resting (no agent process running), the composer could show an effort control the running chat does not have: "Medium" on chats whose model the model list does not include, or a bare "Effort" on main. It disappeared once the agent started, and a pick made in it could not be seen afterwards. #24235 rests a Claude chat after every Stop, so without this fix the stray control would appear after each Stop. What changes for you. A resting chat shows exactly the controls the running chat shows. A normal model keeps its effort control even before Orca has fetched the model list. Fixed during review:
Verified. Live on an M4 Air with real Claude (screenshots in the description): main shows a phantom "Effort" control at rest; #24235 without this PR shows "Medium" after a Stop; with this PR the resting composer matches the running one, both after a Stop and after a relaunch. Tests written to fail first; 254 related test files pass; typecheck clean. Not verified. The idle sweep's own rest (it needs about 35 minutes idle; a quit and relaunch was used instead), Windows, Linux, SSH hosts and the phone app. Deferred. Minor: a saved resolved model id can add a second row beside its alias, only when the running agent also has no model list (live and rest still agree). |
 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| const listed = | ||
| catalog.origin === 'unknown' ? restingFallbackModels(record.provider) : catalog.models |
There was a problem hiding this comment.
sonnet as the default even though the response leaves the current model empty. On a first resting read, the client therefore shows sonnet as selected without knowing what the account runs, and a user can explicitly pick and persist that guess. The new no-pick test applies a live answer first, so it does not cover this case.
Knowledge Base Used: Agent workspaces and sessions
|
Navigate logical layers of code changes, visualize relationships, and explore their blast radius. 🧰 Additional context used📚 Code guidelines (1)No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (5)
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 4 remain after this review. 📝 WalkthroughWalkthroughClaude and Codex catalog construction now use a shared helper to include an unlisted current model. At-rest reads use Claude’s built-in models when the catalog is unknown, but do not infer a current model from those fallback models. Tests compare live and at-rest options for unlisted models, missing catalogs, and saved picks. Priority: ⬇️ Low Merge Risk: ⚪ Minimal · up to This change makes resting native chats list models the same way running chats do, so the stray effort picker should no longer appear. No merge-blocking risk was found. Windows, Linux, SSH hosts, and the phone app were not verified by the author. Security Architecture ReviewSecurity architecture risk: 🔵 Low · up to The change adjusts displayed options without adding write or execution authority. No new security issue was established, but access-control and downstream coverage remain incomplete. Retained concerns Security review detailsSecurity Blast Radius
Trust Boundaries and Controls
Resilience and Maintainability Implications
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches 💡 1📝 Generate docstrings 💡
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![pullfrog[bot]](https://avatars.githubusercontent.com/in/1768019?s=60&v=4){kind=small}
There was a problem hiding this comment.
Important
A resting Claude chat with no saved model and no host catalog now renders the built-in default ("Sonnet") as its selected model, and an at-rest effort change can persist that guess as the model for the next launch. The fix's own guard (leaving current.model empty) is defeated by the fallback list carrying isDefault.
Reviewed changes
- Shared model-list rule: new
structuredAgentSessionOptionModelsappends the current model to a catalog when the catalog does not list it, with no effort levels; the Claude live adapter, the Codex live composer and the at-rest host answer all call it, replacing three copies of the rule. - At-rest Claude fallback: with no host catalog, the resting answer now lists Claude's built-in seed models instead of
[], and names a default only from a real catalog. - Tests: five new cases in
claude-structured-effort-default-at-rest.test.tscovering an unlisted model at rest vs live, the no-catalog fallback for both providers, and the no-pick case.
DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏
| function restingFallbackModels( | ||
| provider: AgentSessionRecord['provider'] | ||
| ): AgentSessionModelOption[] | null { | ||
| return provider === 'claude' ? claudeFallbackModelOptions() : null |
There was a problem hiding this comment.
claudeFallbackModelOptions() keeps isDefault: true on the seed's default row. The at-rest answer sends this list with current.model: '', and the client's structuredAgentSessionOptionCatalog unconditionally sets defaultModelIsCliDefault: true, so the picker names Sonnet even though the host named no model — exactly the guess the comment above says to avoid. Before this PR the no-catalog list was empty, so nothing was named, and a running agent always reports its real model so it never falls back to the seed default either.
Technical details
# No-catalog Claude fallback names a guessed default
## Affected sites
- `src/main/native-chat/agent-session-wire/structured-agent-session-options-read.ts:34` — fallback rows retain `isDefault` from `seedModels()`.
- `src/main/claude/claude-structured-session-options.ts:246` — `claudeFallbackModelOptions` wires `seedModels()` including `isDefault`.
- `src/shared/structured-agent-session-options.ts:89` — client `structuredAgentSessionOptionCatalog` sets `defaultModelIsCliDefault: true` unconditionally.
- `src/shared/native-chat-session-option-snapshot.ts:190` — `cliDefaultModelId` then returns `models.find((m) => m.isDefault)?.id`.
## Required outcome
- A no-catalog, no-saved-pick resting answer must not render a selected model, matching both the pre-PR behavior and a running agent (which reports its actual model).
## Suggested approach (optional)
- Strip `isDefault` from the no-catalog fallback, or gate the client's `defaultModelIsCliDefault` on a non-empty `current.model`.
- `claude-structured-effort-default-at-rest.test.ts:280` asserts the resting and live fallback lists are equal, so stripping only the resting side needs that expectation reconciled.
## Open questions for the human (optional)
- Is a record with no saved model actually reachable at rest (e.g. a first start that never proved)? The added no-pick test treats it as reachable; if it is not, the origin guard here is dead code. The new test only checks `resting.current.model === ''` and applies a prior live read first, so it does not cover the rendered picker.
What this fixes, and where
The composer's option pickers in a resting native Claude or Codex chat, in two cases: the chat's current model is one the model catalog does not list (for example a newer or pinned model id), or the host has no model catalog for the account yet.
A chat rests after the idle sweep today, and after every Stop once #24235 lands. A resting chat on a listed model whose catalog the host holds, PTY sessions, and running chats are unaffected.
ELI5
When a native chat has no live agent process (it is "resting"), the composer can show an extra effort picker ("Medium") that the same chat does not show while the agent is running.
The extra picker is a stray: Orca adds it only because the resting answer leaves the chat's model off the model list. It goes away again as soon as the agent starts.
This PR makes the resting answer list the chat's model exactly the way the running agent does, so the extra picker never appears.
This must land before #24235. That PR rests a Claude chat on every Stop, so without this fix the stray picker would show up right after every Stop.
What Changed
Before: a resting chat on an unlisted model showed an effort picker ("Medium") next to the model. Once the agent started again, the picker disappeared. A pick made in it was saved, but the user could no longer see or change it while the agent ran.
After: a resting chat shows the same pickers as the running chat. There is no effort picker for a model whose effort levels nobody reported, whether or not an agent is running. A listed model keeps its effort picker even before the host has a catalog for the account.
Mechanism:
structuredAgentSessionOptionModels. It returns the catalog's rows plus the current model when the catalog does not list it, with no effort levels.opuskeeps its effort picker.Wire and versions: this is host-only and the options message is unchanged.
Why
The rule this enforces: whether a model has an effort control should depend on the model and the catalog, not on whether an agent process happens to be alive.
Why one shared function, not a patch to the resting path: the bug came from two copies of the same rule drifting apart. A shared function removes the drift instead of adding a third copy.
What I did not do:
Not in scope: why a given Claude install reports its running model as an id its own catalog does not list. That is a separate question.
Differences from the common pattern
The mechanism matches the common pattern: the effort control is derived from the catalog and the selected model, the same way whether or not a session is live.
On a model the catalog does not list, the common pattern splits: one shows no effort control, the other shows the default model's levels. Intended: this PR keeps what Orca's running agents already do (no effort control), so the composer does not change across a Stop. Showing another model's levels would offer choices nothing says the chat's model accepts.
With no catalog at all, the common pattern falls back to a built-in list that has effort levels. This PR does the same for Claude, with the list a running Claude agent already falls back to. For Codex the host has no built-in list, so the client fills the current model from its own unknown-model defaults, unchanged from before this PR.
Linked Issue
Found while reviewing #24235 (Claude Stop ends the agent process). Needed before #24235 lands.
Visual Proof
Live Claude on an M4 Air, a chat whose model the catalog does not list (shown as the raw id), composer at rest (no Claude process for the chat). Builds: 2cedb9d; the head differs only in the no-catalog default model.
For comparison, the same chat while Claude runs shows the same controls as at rest:
Testing
Red first:
src/main/claude/claude-structured-effort-default-at-rest.test.ts: a running Claude agent whose model the catalog does not list, compared with the same chat at rest. The row for that model must be identical, and the client must show no effort picker.opuslists exactly what a running agent whose CLI lists nothing lists, and keeps its effort picker at the saved level.opus), not the built-in list's default (sonnet). It failed before this fix withexpected 'sonnet' to be ''.opus, with no effort levels, and the Codex answer listedgpt-5.5with none.Kept green: the existing at-rest effort cases for a listed model (
opus[1m]showing its learned default, or the user's pick).Ablations:
Runs (explicit files):
Lint: oxlint, the changed-code quality gate and the import-cycle check are clean.
I manually tested these changes locally (live on an M4 Air, see Visual Proof)
Automated tests added/updated, or explained why not below
AI Disclosure
Review
Agent skill upstream boundary
docs/reference/agent-skill-sharing-upstream-boundary.mdand copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.Notes
Checklist
N/Awith reasonpnpm lint,pnpm typecheck,pnpm test, andpnpm buildpass (or CI will cover; local preferred)